Self-consistency, Extract and Rectify: Knowledge Graph Enhance Large Language Model for Electric Power Question Answering

Electric power artificial intelligence has rapidly advanced in recent years, encompassing safety detection, assistant decision-making, and optimal scheduling. With the rise of Large Language Models (LLMs), knowledge-based AI is becoming increasingly prevalent across various domains. However, in the field of electric power, most of the knowledge-based AI is centered on Knowledge Graph (KG) techniques, while less research has been done on power LLMs. In this paper, we are inspired by Self-Consistency (SC) and propose a Self-Consistency, Extraction and Rectify framework—SCER, for the usage of KG-enhanced LLM in power operations and maintenance (O&M) question answering scenarios. Specifically, we transfer the SC from the general-purpose domain into the power domain and replace the original model with a Chinese sentence representation model to make it more localized. We design an Extract Mechanism to generate evidence chains through multiple random walks on the POMKG and a Rectify Mechanism to correct the score of the generated rationales. Extensive experiments and specific case studies on the POMQA dataset demonstrate the effectiveness of our proposed SCER for SC transfer and improvement in the power field.